LLM API Pricing History: Every Price Drop Since 2023

The best-known charts of falling LLM prices are essays. a16z's "LLMflation" piece from November 2024 and Epoch AI's analysis of inference price trends are both careful and both widely linked, and both describe the market on the day they were written. If you want the number for this month, you are on your own.

This page is that measurement kept running. Every day a script reads three public sources, works out the cheapest price for three fixed levels of model capability, and rebuilds the tables below. The history goes back to October 2023, rebuilt month by month from the git history of the open-source LiteLLM price file.

I run Gravity, where every agent run is paid for in tokens, so I check these prices more than most people. Writing the dates down was the only way to answer the question I kept getting asked: how fast is this actually falling?

Updated 4 October 2026. Sources for this update: LiteLLM price file commit a66adb4, OpenRouter models API (466 models listed), LMArena leaderboard of 2 October 2026.

Price per million tokens by capability tier

Each line below holds capability fixed and asks what the cheapest way to buy it costs. A model counts toward a tier when its LMArena rating is at or above the tier's floor model: the original GPT-4 from March 2023, GPT-4o from May 2024, or Gemini 2.5 Pro from 2025. The price is the cheapest blended rate among those models, counting the labs' own APIs and six large inference hosts.

Month by month

The first snapshot of every month, newest first, with the model that set the floor and who served it.

Blended price = (3 × input + output) ÷ 4, in US dollars per million tokens. Monthly rows read the LiteLLM price file as it stood at 00:00 UTC on the 1st of the month (from its git history); the top row reads today's file. Tier membership uses the LMArena leaderboard of 2 October 2026.

The moves that mattered

The GPT-4 class line is the longest, and its steps are worth reading one at a time, because each one is a different kind of price drop.

• October and November 2023: $37.50 blended. GPT-4 itself was the only model in its class, at $30 input and $60 output per million tokens.
• December 2023: $15. GPT-4 Turbo (gpt-4-1106-preview) arrived at $10 input and $30 output. The same lab, a newer model, less than half the price.
• June 2024: $7.50. GPT-4o at $5 and $15. It also set the floor for the GPT-4o class, which started that month.
• August 2024: $0.26. gpt-4o-mini at $0.15 input and $0.60 output. The floor fell by more than 95% in one month, because a small model cleared a bar that only frontier models had cleared before.
• October 2024 onward: about $0.13. Google's Gemini 1.5 Flash-002 and then Gemini 2.0 Flash-Lite, at $0.075 input and $0.30 output.
• September 2025 onward: a few cents. Small open-weight models (Gemma 3 4B, then gpt-oss-20b) served by DeepInfra. From here the floor belongs to hosts competing on open models, not to the labs.

The Gemini 2.5 Pro class shows the same pattern, faster. It started in July 2025 at $3.44 blended (Gemini 2.5 Pro itself, $1.25 input and $10 output) and by September 2026 an open-weight model on a host had taken the floor at about 15 cents. The frontier gets cheaper slowly; the level the frontier reached a year ago gets cheaper very fast.

If you are pricing an AI product, that gap is the useful part. I wrote up what it means for agents specifically in what it costs to run an AI agent around the clock and AI agent cost optimization.

Current API prices from 11 labs

The five newest paid text models from each lab, with input, output and cached-input prices per million tokens and the context window. This is the table to check for OpenAI API pricing, Claude API pricing, Gemini API pricing or DeepSeek API pricing on a given day. Prices are the lab's own list price wherever the model matches the lab's own entry in the LiteLLM file.

The 5 newest paid text models per lab, by the date OpenRouter listed them, read on 4 October 2026. Prices in US dollars per million tokens are the lab's own list price from the LiteLLM file where the model matches one of the lab's own entries; rows marked "OpenRouter price" carry OpenRouter's listed price instead, which for open-weight models can be a third-party host's. Context is OpenRouter's figure. "Last list-price change seen" compares the lab's own entries in the LiteLLM file between snapshots: month by month before 4 October 2026, day by day after.

List price is the starting point, not the bill. Batch APIs, prompt caching, long-context surcharges and reasoning tokens all move the real number, and the AI agent pricing index shows how platforms built on these models turn token prices into plan prices.

Price change log

Dated list-price changes on the labs' own models, newest first. Monthly windows before 4 October 2026 come from the git history; from that date on, each line is a single day.

1. between 1 September 2026 and 1 October 2026: Mistral, mistral/mistral-medium, input $2.70 to $1.50, output $8.10 to $7.50 per million tokens (LiteLLM price file)
2. between 1 September 2026 and 1 October 2026: DeepSeek, deepseek-v4-flash, input $0.44 to $0.30, output $1.32 to $1.20 per million tokens (LiteLLM price file)
3. between 1 September 2026 and 1 October 2026: DeepSeek, deepseek-v4-flash-vision-exp, input $0.44 to $0.30, output $1.32 to $1.20 per million tokens (LiteLLM price file)
4. between 1 August 2026 and 1 September 2026: Google, gemini/gemini-3.6-flash, input $1.50 to $0.75, output $7.50 to $3.75 per million tokens (LiteLLM price file)
5. between 1 August 2026 and 1 September 2026: OpenAI, gpt-5.6, input $5.00 to $4.00, output $30.00 to $20.00 per million tokens (LiteLLM price file)
6. between 1 August 2026 and 1 September 2026: OpenAI, gpt-5.6-sol, input $5.00 to $4.00, output $30.00 to $20.00 per million tokens (LiteLLM price file)
7. between 1 August 2026 and 1 September 2026: xAI, xai/grok-3, input $3.00 to $1.25, output $15.00 to $2.50 per million tokens (LiteLLM price file)
8. between 1 August 2026 and 1 September 2026: xAI, xai/grok-3-beta, input $3.00 to $1.25, output $15.00 to $2.50 per million tokens (LiteLLM price file)
9. between 1 August 2026 and 1 September 2026: xAI, xai/grok-3-fast-beta, input $5.00 to $1.25, output $25.00 to $2.50 per million tokens (LiteLLM price file)
10. between 1 August 2026 and 1 September 2026: xAI, xai/grok-3-mini, input $0.30 to $1.25, output $0.50 to $2.50 per million tokens (LiteLLM price file)
11. between 1 August 2026 and 1 September 2026: xAI, xai/grok-3-mini-beta, input $0.30 to $1.25, output $0.50 to $2.50 per million tokens (LiteLLM price file)
12. between 1 August 2026 and 1 September 2026: xAI, xai/grok-3-mini-fast, input $0.60 to $1.25, output $4.00 to $2.50 per million tokens (LiteLLM price file)
13. between 1 August 2026 and 1 September 2026: xAI, xai/grok-3-mini-fast-beta, input $0.60 to $1.25, output $4.00 to $2.50 per million tokens (LiteLLM price file)
14. between 1 August 2026 and 1 September 2026: xAI, xai/grok-4, input $3.00 to $1.25, output $15.00 to $2.50 per million tokens (LiteLLM price file)
15. between 1 August 2026 and 1 September 2026: xAI, xai/grok-4-fast-reasoning, input $0.20 to $1.25, output $0.50 to $2.50 per million tokens (LiteLLM price file)
16. between 1 August 2026 and 1 September 2026: xAI, xai/grok-4-fast-non-reasoning, input $0.20 to $1.25, output $0.50 to $2.50 per million tokens (LiteLLM price file)
17. between 1 August 2026 and 1 September 2026: xAI, xai/grok-4-0709, input $3.00 to $1.25, output $15.00 to $2.50 per million tokens (LiteLLM price file)
18. between 1 August 2026 and 1 September 2026: xAI, xai/grok-4-1-fast, input $0.20 to $1.25, output $0.50 to $2.50 per million tokens (LiteLLM price file)
19. between 1 August 2026 and 1 September 2026: xAI, xai/grok-4-1-fast-reasoning, input $0.20 to $1.25, output $0.50 to $2.50 per million tokens (LiteLLM price file)
20. between 1 August 2026 and 1 September 2026: xAI, xai/grok-4-1-fast-non-reasoning, input $0.20 to $1.25, output $0.50 to $2.50 per million tokens (LiteLLM price file)
21. between 1 August 2026 and 1 September 2026: xAI, xai/grok-4.20-beta-0309-reasoning, input $2.00 to $1.25, output $6.00 to $2.50 per million tokens (LiteLLM price file)
22. between 1 August 2026 and 1 September 2026: xAI, xai/grok-4.20-0309-reasoning, input $2.00 to $1.25, output $6.00 to $2.50 per million tokens (LiteLLM price file)
23. between 1 August 2026 and 1 September 2026: xAI, xai/grok-4.20-beta-0309-non-reasoning, input $2.00 to $1.25, output $6.00 to $2.50 per million tokens (LiteLLM price file)
24. between 1 August 2026 and 1 September 2026: xAI, xai/grok-code-fast, input $0.20 to $1.00, output $1.50 to $2.00 per million tokens (LiteLLM price file)
25. between 1 August 2026 and 1 September 2026: xAI, xai/grok-code-fast-1, input $0.20 to $1.00, output $1.50 to $2.00 per million tokens (LiteLLM price file)
26. between 1 August 2026 and 1 September 2026: xAI, xai/grok-code-fast-1-0825, input $0.20 to $1.00, output $1.50 to $2.00 per million tokens (LiteLLM price file)
27. between 1 August 2026 and 1 September 2026: DeepSeek, deepseek-v4-flash, input $0.14 to $0.44, output $0.28 to $1.32 per million tokens (LiteLLM price file)
28. between 1 August 2026 and 1 September 2026: DeepSeek, deepseek-v4-pro, input $0.435 to $1.32, output $0.87 to $3.96 per million tokens (LiteLLM price file)
29. between 1 July 2026 and 1 August 2026: Anthropic, claude-sonnet-5, input $3.00 to $2.00, output $15.00 to $10.00 per million tokens (LiteLLM price file)
30. between 1 June 2026 and 1 July 2026: Cohere, command-r7b-12-2024, input $0.15 to $0.0375, output $0.0375 to $0.15 per million tokens (LiteLLM price file)
31. between 1 January 2026 and 1 February 2026: DeepSeek, deepseek-chat, input $0.60 to $0.28, output $1.70 to $0.42 per million tokens (LiteLLM price file)
32. between 1 January 2026 and 1 February 2026: DeepSeek, deepseek-reasoner, input $0.60 to $0.28, output $1.70 to $0.42 per million tokens (LiteLLM price file)
33. between 1 January 2026 and 1 February 2026: DeepSeek, deepseek/deepseek-chat, input $0.27 to $0.28, output $1.10 to $0.42 per million tokens (LiteLLM price file)
34. between 1 January 2026 and 1 February 2026: DeepSeek, deepseek/deepseek-reasoner, input $0.55 to $0.28, output $2.19 to $0.42 per million tokens (LiteLLM price file)
35. between 1 December 2025 and 1 January 2026: Moonshot AI, moonshot/kimi-thinking-preview, output $30.00 to $2.50 per million tokens (LiteLLM price file)
36. between 1 September 2025 and 1 October 2025: OpenAI, gpt-3.5-turbo, input $1.50 to $0.50, output $2.00 to $1.50 per million tokens (LiteLLM price file)
37. between 1 June 2025 and 1 July 2025: OpenAI, o3, input $10.00 to $2.00, output $40.00 to $8.00 per million tokens (LiteLLM price file)
38. between 1 June 2025 and 1 July 2025: OpenAI, o3-2025-04-16, input $10.00 to $2.00, output $40.00 to $8.00 per million tokens (LiteLLM price file)
39. between 1 June 2025 and 1 July 2025: Google, gemini/gemini-2.5-flash-preview-05-20, input $0.15 to $0.30, output $0.60 to $2.50 per million tokens (LiteLLM price file)
40. between 1 March 2025 and 1 April 2025: Mistral, mistral/mistral-small, input $1.00 to $0.10, output $3.00 to $0.30 per million tokens (LiteLLM price file)

The 40 most recent list-price changes on the 11 labs' own models, newest first. A change in the LiteLLM file can also be a correction to that file, which is why each line names its source. New models are not price changes and do not appear here.

Methodology

Sources. Prices come from LiteLLM's model_prices_and_context_window.json, an MIT-licensed file that the LiteLLM project keeps in step with provider pricing pages. Its git history supplies one snapshot per month: the file as it stood at 00:00 UTC on the 1st, from October 2023 (the first month boundary after the file appeared) to October 2026. Today's prices come from the same file on its main branch. The per-lab table lists models from the OpenRouter models API, which also supplies the context window. Capability ratings come from the LMArena leaderboard dataset (CC BY 4.0).

Blended price. One input-to-output weighting, three parts input to one part output, turns two prices into one comparable number. It is the convention Artificial Analysis uses, and it suits chat-style workloads; output-heavy work such as long generation or reasoning will cost more than the blend suggests.

Who counts. The tier floors count the 11 labs' own APIs (OpenAI, Anthropic, Google's Gemini API, xAI, DeepSeek, Mistral, Alibaba Cloud, Moonshot AI, Z.ai, MiniMax, Cohere) and six inference hosts with public per-token prices: Together AI, Fireworks AI, Groq, DeepInfra, Cerebras and SambaNova. Resellers, cloud marketplaces (Bedrock, Vertex, Azure) and entries priced at zero are left out. Standard pay-as-you-go rates only: no batch, priority or long-context tiers.

How the tiers are defined

A tier is a rating band on LMArena's text leaderboard (style control, overall category). A model is in the GPT-4 class when its rating is at or above gpt-4-0314's, in the GPT-4o class at or above gpt-4o-2024-05-13's, and in the Gemini 2.5 Pro class at or above gemini-2.5-pro's. LMArena ratings come from human preference votes, so a tier measures what people preferred in blind comparisons, which is not the same as coding or reasoning benchmarks. I chose it because it is public, covers hundreds of models from every major lab, and is published under an open licence.

LMArena model names are matched to LiteLLM entries by name, after removing provider prefixes and suffixes such as "-instruct" or a reasoning-effort label. Unmatched models simply do not count, which can only make a floor higher than the true cheapest price, never lower. Ratings refresh monthly; a refresh can move models in or out of a tier for later snapshots, and earlier snapshots keep the result they were computed with.

Known gaps

• Google before October 2024. Until then the LiteLLM file carried no per-token price for Gemini API models (zero or per-character Vertex rates), so Google's models enter the series in October 2024. Gemini 1.5 Flash was available earlier, so the GPT-4 class floor for mid-2024 may be overstated.
• Corrections look like price changes. When the LiteLLM maintainers fix a wrong price, the change log records it like a provider move. Changes larger than 20 times in either direction are treated as unit fixes and skipped.
• OpenRouter prices for open models. Where a model has no first-party entry in LiteLLM, the per-lab table shows OpenRouter's listed price, which for open-weight models can come from a third-party host.
• Time-of-day pricing. DeepSeek's own pricing page lists peak and off-peak rates, with off-peak at half the peak price as read on 4 October 2026. The LiteLLM file, and so this page, carries the peak rate.
• Quality is not identical inside a tier. A model at the tier floor matched the anchor on one leaderboard. Hosts also differ on quantization, speed, rate limits and data retention.

Free usage is a separate question with its own tracker: free LLM API tiers, verified daily. For what agent platforms charge on top of these tokens, see the cheapest AI agent platforms and AI agent pricing trends in 2026. Gravity's own plans are on the pricing page: Basic is free with one agent, and Pro is $5 a month during the alpha.

FAQ

Sources